問題文
An internal agent reads issue tickets submitted by external customers and can call a tool that closes tickets. A customer submits a ticket whose body contains text directing the assistant to close all open tickets for their account. Which two measures address this? (Select two.)
選択肢
- Raise the temperature so the agent's behavior is harder to predict and therefore harder to manipulate.
- Pass the ticket body as clearly delimited untrusted content that is explicitly not to be followed as instructions, keeping the operating rules in the system parameter.
- Restrict the close-ticket tool so it can only act on the single ticket under review, and require an explicit approval step for any bulk action.
- Add a line to the system prompt asking customers not to include instructions in ticket bodies.
- Switch to a more capable model grade, since stronger models follow the true instructions and ignore injected ones.